license: mit
InstructProtein
InstructProtein is the first large generative language model exploring the feasibility of bidirectional generation between human and protein language. It is based on OPT-1.3B architecture with two-step training approach: It initiates with pre-training on protein and natural language corpora, followed by fine-tuning with the established protein knowledge instruction dataset. Through further instruction tuning, InstructProtein outperforms larger general-purpose foundation models on protein understanding and design tasks.
Limitations
The current model, developed through instruction tuning using knowledge instruction dataset, serves as a preliminary example. Despite its initial success in controlled environments, it lacks the robustness to manage complex, real-world, production-level tasks.
Reference
For more information, please take a look at our paper and repository.